U.S. patent number 7,085,749 [Application Number 10/155,478] was granted by the patent office on 2006-08-01 for pulse signal circuit, parallel processing circuit, pattern recognition system, and image input system.
This patent grant is currently assigned to Canon Kabushiki Kaisha. Invention is credited to Masakazu Matsugu.
United States Patent |
7,085,749 |
Matsugu |
August 1, 2006 |
**Please see images for:
( Certificate of Correction ) ** |
Pulse signal circuit, parallel processing circuit, pattern
recognition system, and image input system
Abstract
A synaptic connection element for connecting neuron elements
inputs a plurality of pulsed signals from different neuron elements
N1 through N4, effects a common modulation (time window integration
or pulse phase/width modulation) on a plurality of predetermined
signals among the plurality of pulse signals, and outputs the
modulated pulse signals to different signal lines to a neuron
element M1. A neural network for representing and processing
pattern information by the pulse modulation is thereby downsized in
scale.
Inventors: |
Matsugu; Masakazu (Chiba,
JP) |
Assignee: |
Canon Kabushiki Kaisha (Tokyo,
JP)
|
Family
ID: |
26616078 |
Appl.
No.: |
10/155,478 |
Filed: |
May 28, 2002 |
Prior Publication Data
|
|
|
|
Document
Identifier |
Publication Date |
|
US 20030004907 A1 |
Jan 2, 2003 |
|
Foreign Application Priority Data
|
|
|
|
|
May 31, 2001 [JP] |
|
|
2001/164508 |
May 28, 2002 [JP] |
|
|
2002/153997 |
|
Current U.S.
Class: |
706/20; 706/33;
706/35; 327/355 |
Current CPC
Class: |
G06K
9/00973 (20130101); G06N 3/063 (20130101); G06N
3/049 (20130101); G06K 9/4628 (20130101) |
Current International
Class: |
G06N
3/04 (20060101); G06F 15/18 (20060101); G06G
7/161 (20060101) |
Field of
Search: |
;706/20,22 |
References Cited
[Referenced By]
U.S. Patent Documents
Foreign Patent Documents
|
|
|
|
|
|
|
60-712 |
|
Jan 1985 |
|
JP |
|
5-37317 |
|
Feb 1993 |
|
JP |
|
7-262157 |
|
Oct 1995 |
|
JP |
|
7-334478 |
|
Dec 1995 |
|
JP |
|
8-153148 |
|
Jun 1996 |
|
JP |
|
2741793 |
|
Jan 1998 |
|
JP |
|
10-327054 |
|
Dec 1998 |
|
JP |
|
2879670 |
|
Jan 1999 |
|
JP |
|
Other References
Ota et al., "Analog Implementation of Pulse-Coupled Neural
Networks", IEE, Transactions On Neural Networks, vol. 10, No. 3,
pp. 539-544. cited by examiner .
Tomberg et al., Pulse-Density Modulation Technique in VLSI
Implementations of Neural Network Algorithms, 1990, IEE, Journal Of
Solid-State Circuits, vol. 25, No. 5, pp. 1277-1286. cited by
examiner .
Sirovich, et al., Mar. 1987, "Low-Dimensional Procedure for
Characterization of Human Faces," J. Opt. Soc. Am. [A], vol. 4, No.
3, pp. 519-524. cited by other .
Lades, et al., Mar. 1993, "Distortion Invariant Object Recognition
in the Dynamic Link Architecture, " IEEE Trans. On Computers, vol.
42, pp. 300-311. cited by other .
Seibert, et al., 1992, "Learning and Recognizing 3D Objects from
Multiple Views in a Neural System," In Neural Netword for
Perception, vol. 1, Human and Machine Perception, Edidted by H.
Wechsler, Academic Press, pp. 426-444. cited by other .
Fukushima, et al., 1982, "Neocognitron: A New Algorithm for Pattern
Recognition Tolerant f Deformation and Shifts in Position, "
Pattern Recognition, vol. 15, No. 6, pp. 455-469. cited by other
.
Anderson, et al., 1995, "Routing Networks in Visual Cortex, " In
Handbook of Brain Theory and Neural Networks, MIT Press, pp.
823-826. cited by other .
Murray, et al., Mar. 1991, "Pulse-Stream VLSI Neural Network Mixing
Analog and Digital Techniques," IEEE Trans. On Neural Networks,
vol. 2, No. 2, pp. 193-204. cited by other .
Broussard, et al., May 1999, "Physiologically Motivated Image
Fusion for Object Detecti n Using a Pulse Coupled Neural Network, "
IEEE Trans. on Neural Networks, vol. 10, pp. 554-563. cited by
other .
Eckhorn, et al., Fall 1990, "Feature Linking via Synchronization
Among Distributed Assemblies: Simulations of Results from Cat
Visual Cortex," Neural Computation, vol. 2, pp. 293-307. cited by
other .
Lazzaro, et al., 1993, "Silicon Auditory Processors as Computer
Peripherals," in Advances in Neural Infor. Proc. System 5, edited
by Hanson, et al. cited by other .
LeCur, et al., Jul. 1995, "Convolutional Networks for Images
Speech, and Time Series", In Handbook of Brain Theory and Neural
Networkds, MIT Press, pp. 225-258. cited by other .
Daugman, Jul. 1988, "Complete Discrete 2-D Gabor transforms by
Neural Networks for Image Analysis and Compression, " IEEE Trans.
on Acoustics, Speed, and Signal Processing, vol. 36, No. 7, pp.
1169-1179. cited by other.
|
Primary Examiner: Knight; Anthony
Assistant Examiner: Brown, Jr.; Nathan H.
Attorney, Agent or Firm: Fitzpatrick, Cella, Harper &
Scinto
Claims
What is claimed is:
1. A pulse signal processing circuit comprising: a plurality of
arithmetic elements for outputting pulse signals; and at least one
modulation circuit each for receiving a plurality of pulse signals
from different arithmetic elements, modulating in common a
plurality of predetermined pulse signals among the plurality of
pulse signals, and outputting the modulated pulse signals in branch
to different signal lines, respectively, wherein said at least one
modulation circuit each gives a time-differing modulation
characteristic with respect to a process of the same input
data.
2. A pulse signal processing circuit according to claim 1, wherein
said at least one modulation circuit each is a delay circuit for
giving a predetermined delay to the plurality of predetermined
pulse signals among the plurality of pulse signals.
3. A parallel processing circuit comprising: a plurality of neuron
elements; and synaptic connection elements for connecting said
neuron elements, wherein said synaptic connection elements are
constructed by use of said pulse signal processing circuit claimed
in claim 1.
4. A parallel processing circuit comprising: a plurality of neuron
elements; and synaptic connection elements for connecting said
neuron elements, which comprise, a plurality of arithmetic elements
for outputting pulse signals, and at least one modulation circuit
each for receiving a plurality of pulse signals from different
arithmetic elements, modulating in common a plurality of
predetermined pulse signals among the plurality of pulse signals,
and outputting the modulated pulse signals in branch to different
signal lines, respectively, wherein said synaptic connection
element shares a plurality of synaptic connections as one circuit
on the basis of a distribution symmetry of synaptic connections to
said predetermined ones of said neuron elements.
5. A pattern recognition system comprising: data input means for
inputting data of a predetermined dimension; a plurality of data
processing modules having feature detection layers for detecting a
plurality of features; and output means for outputting a result of
a pattern recognition, wherein said data processing module includes
a plurality of arithmetic elements connected to each other by
predetermined synaptic connection means, each of said arithmetic
elements outputs a pulse signal at a frequency or timing
corresponding to an arrival time pattern of a plurality of pulses
within a predetermined time window, said output means outputs,
based on the outputs of said plurality of arithmetic elements, a
result of detecting or recognizing a predetermined pattern, said
synaptic connection means includes at least one modulation circuit
each for receiving the plurality of pulse signals from said
different arithmetic elements, effecting a predetermined common
modulation on a plurality of predetermined pulse signals among the
plurality of pulse signals, and outputting in branch the modulated
pulse signals to different signal lines, respectively, and said at
least one modulation circuit each gives a time-differing modulation
characteristic with respect to a process of the same input
data.
6. A pattern recognition system according to claim 5, wherein said
at least one modulation circuit each is a delay circuit for giving
a predetermined delay to the plurality of predetermined pulse
signals among the plurality of pulse signals.
7. A pattern recognition system according to claim 5, wherein said
at least one modulation circuit each is a pulse width modulation
circuit for effecting a predetermined pulse width modulation on the
plurality of predetermined pulse signals among the plurality of
pulse signals.
8. A pattern recognition system comprising: a modulation circuit
for receiving a plurality of pulse signals from different
arithmetic elements and giving a predetermined delay to each pulse
signal; and a branch circuit for outputting in branch the modulated
pulse signals in a predetermined sequence to different signal
lines, respectively, in a way that gives a predetermined delay to
each pulse signal, wherein said modulation circuit gives a
time-differing modulation characteristic with respect to a process
of the same input data.
9. A pulse signal processing circuit comprising: a parallel
modulation circuit for receiving a plurality of pulse signals from
different arithmetic elements and effecting a predetermined
modulation in parallel on a plurality of predetermined pulse
signals among the plurality of pulse signals; and integration means
for integrating outputs of said parallel modulation circuit,
wherein said parallel modulation circuit includes a plurality of
time window integration circuits for effecting a predetermined
weighted time window integration with respect to a plurality of
predetermined pulse signals among the modulated pulse signals, each
of said arithmetic elements output a predetermined pulse signal on
the basis of a signal from said integration means, and said time
window integration circuits multiply the pulse signals by a
predetermined weighting coefficient set for every pulse signal
inputted in parallel to said parallel modulation circuit and
configure receptive fields of said arithmetic elements.
10. A pulse signal processing circuit according to claim 9, wherein
said parallel modulation circuit is a pulse width modulation
circuit for effecting a predetermined pulse width modulation on the
plurality of predetermined pulse signals among the plurality of
pulse signals.
11. A pulse signal processing circuit according to claim 9, wherein
outputs of said arithmetic elements are, after being stored in a
predetermined memory, outputted in branch to other arithmetic
elements.
12. A pulse signal processing circuit according to claim 9, further
comprising a weighting coefficient signal generation circuit,
wherein said time window integration circuits multiply the input
pulse signals by a weighting coefficient signal transmitted from
said weighting coefficient signal generation circuit.
13. An image input system executing a predetermined image input by
use of said pulse signal processing circuit claimed in claim 9.
Description
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates to a pulse signal circuit and further
to a parallel processing circuit and a pattern recognition system
that use this pulse signal circuit, and an image input system for
detecting a specified object etc. by use of the pattern recognition
system.
2. Related Background Art
Image and voice recognition implementation systems have hitherto
been roughly classified into such a type that a recognition
processing algorithm specialized for a specified recognition object
is sequentially operated and executed as computer software, and a
type in which the same algorithm is executed by a dedicated
parallel image processor (such as an SIMD (Single Instruction
Multiple Data) processor, an MIMD (Multiple Instruction
stream/Multiple Data stream) processor and so on).
Typical examples are given below as exemplifying the image
recognition algorithm. At first, the following is methods involving
calculating a feature amount relative to a similarity to a
recognition object model. One method is a method for representing
recognition object model data as a template model, calculating a
similarity by template matching etc with an input image (or a
feature vector thereof) and calculating a high-order correlation
coefficient. Another method is a method (Sirovich, et al., 1987,
Low-dimensional procedure for the characterization of human faces,
J. Opt. Soc. Am. [A], vol. 3, pp. 519 524) for mapping an input
pattern to an intrinsic image function space obtained by analyzing
primary components of an object model image, and calculating an
intra-feature-space distance from the model. A further method is a
method (Lades et al., 1993, Distortion Invariant Object Recognition
in the Dynamic Link Architecture, IEEE Trans. on Computers, vol.
42, pp. 300 311) for representing a plurality of feature extraction
results (feature vectors) and a spatial arrangement relationship as
graphs, and calculating a similarity based on elastic graph
matching. A still further method is a method (Seibert, et al.,
1992, Learning and recognizing 3D objects from multiple views in a
neural system, in Neural Networks for Perception, vol. 1 Human and
Machine Perception (H. Wechsler Ed.) Academic Press, pp. 427 444)
for obtaining position-, rotation- and scale-invariable
representations by executing predetermined conversions with respect
to input images and thereafter collating with a model.
The following is exemplifications of a pattern recognition method
based on a neural network model of which a hint is acquired from a
biological information processing system. One exemplification is a
method (Japanese Patent Post-Exam. No. 60-712, Fukushima and
Miyake, 1982, Neocognitron: A new algorithm for pattern recognition
tolerant of deformation and shifts in position, Pattern
Recognition, vol. 15, pp--455 469) for implementing hierarchical
template matching. Another exemplification is a method (Anderson,
et al., 1995, Routing Networks in Visual Cortex, in Handbook of
Brain Theory and Neural Networks (M. Arbib, Ed.), MIT Press, pp.
823 826) for obtaining object-based scale- and position-invariable
representations by dynamic routing neural networks. Other
exemplifications are methods using multi-layer perceptrons, a
radial basis function network and so on.
On the other hand, what is proposed as a scheme for taking an
information processing system based on biological neural networks
with a higher fidelity, is a neural network model circuit (Murray
et al., 1991, Pulse-Stream VLSI Neural Networks Mixing analog and
digital Techniques, IEEE Trans. on Neural Networks, vol. 1.2, pp.
193 204; Japanese Patent Application Laid-Open Nos. 7-262157,
7-334478 and 8-153148, and Japanese Patent Publication No. 2879670)
for transmitting and representing information through on a pulse
train corresponding to an action potential.
Methods for recognizing and detecting a specified object by a
neural network constructed of pulse train generation neurons, are
systems (U.S. Pat. No. 5,664,065, and Broussard, et al., 1999,
Physiologically Motivated Image Fusion for Object Detection using a
Pulse Coupled Neural Network, IEEE Trans. on Neural Networks, vol.
10, pp. 554 563, and so forth) using a pulse coupled neural network
(which will hereinafter be abbreviated to PCNN), to be specific, a
high-order (second-order or higher) model by Echhorn (Eckhorn, et
al., 1990, Feature linking via synchronization among distributed
assembles: simulation of results from cat cortex, Neural
Computation, vol. 2, pp. 293 307) which is based on the premise of
linking inputs and feeding inputs.
Further, a method for relieving a wiring problem in the neural
network is an event-driven oriented method (Address Event
Representation: this will hereinafter be abbreviated to AER)
(Lazzaro, et al., 1993, silicon Auditory Processors as Computer
Peripherals, In Touretzky, D (ed), Advances in Neural Information
Processing Systems 5. San Mateo, Calif.: Morgan Kaufmann
Publishers) for coding addresses of so-called pulse output neurons.
In this case, IDs of pulse train output-sided neurons are coded as
binary addresses, whereby even when output signals from the
different neurons are arranged in time sequence on a common bus,
the input-sided neurons are able to automatically decode the
addresses of the source neurons.
On the other hand, the neural network processor related to Japanese
Patent Publication No. 2741793 schemes to reduce the number of
neurons and to downsize a circuit by configuring multi-layered
feedforward oriented networks in a systolic array architecture.
Each of the prior arts described above, however, still entails, as
problems to a great extent, a difficulty of downsizing both wiring
portions related to inter-neuron connections and a circuit scale of
synaptic connection circuit of which the number is by far larger
than the number of neurons, and a difficulty of applying a layout
of respective components to a general pattern recognition.
SUMMARY OF THE INVENTION
It is therefore a primary object of the present invention to
actualize an equal performance in a smaller circuit scale than in
the prior arts by sharing a synaptic connection circuit.
According to one aspect, the present invention which achieves these
objectives relates to a pulse signal processing circuit comprising
a modulation circuit for inputting a plurality of pulsed signals
from different arithmetic elements and modulating in common a
plurality of predetermined signals among the plurality of pulse
signals, wherein the modulated pulse signals are outputted in
branch to different signal lines, respectively.
According to another aspect, the present invention which achieves
these objectives relates to a pattern recognition system comprising
a data input unit for inputting data of a predetermined dimension,
a plurality of data processing modules having feature detection
layers for detecting a plurality of features, and an output unit
for outputting a result of a pattern recognition, wherein the data
processing module includes a plurality of arithmetic elements
connected to each other by a predetermined synaptic connection
unit, each of the arithmetic element outputs a pulsed signal at a
frequency or timing corresponding to an arrival time pattern of a
plurality of pulses within a predetermined time window, the output
unit outputs, based on the outputs of the plurality of arithmetic
elements, a result of detecting or recognizing a predetermined
pattern, and the synaptic connection unit includes a modulation
circuit for inputting the plurality of pulsed signals from the
different arithmetic elements and effecting a predetermined common
modulation on a plurality of predetermined pulsed signals among the
plurality of pulsed signals, and outputs in branch the modulated
pulse signals to different signal lines.
According to still another aspect, the present invention which
achieves these objectives relates to a pulse signal processing
circuit comprising a modulation circuit for inputting a plurality
of pulsed signals from different arithmetic elements and giving a
predetermined delay to each pulse, and a branch circuit for
outputting in branch the modulated pulse signals in a predetermined
sequence to different signal lines respectively in a way that gives
a predetermined delay to the pulse signal.
According to yet another aspect, the present invention which
achieves these objectives relates to a parallel processing circuit
comprising a plurality of arithmetic elements, arrayed in parallel,
for extracting a different feature pattern category in every
predetermined area with respect to a predetermined sampling
position on input data of a predetermined dimension, wherein each
of the arithmetic elements is connected to other predetermined
arithmetic element through synaptic connection unit, and the
plurality of arithmetic elements for extracting the different
feature pattern category relative to the predetermined position on
the input data, are disposed adjacent to each other.
According to yet another aspect, the present invention which
achieves these objectives relates to a pulse signal processing
circuit comprising a parallel modulation circuit for inputting a
plurality of pulsed signals from different arithmetic elements and
effecting a predetermined modulation in parallel on a plurality of
predetermined signals among the plurality of pulse signals, and an
integration unit for integrating outputs of the parallel modulation
circuit, wherein the parallel modulation circuit includes a
plurality of time window integration circuits for effecting a
predetermined weighted time window integration with respect to a
plurality of predetermined signals among the modulated pulse
signals, and the arithmetic element outputs a predetermined pulse
signal on the basis of a signal from the integration unit.
According to a further aspect, the present invention which achieves
these objectives relates to a pulse signal processing circuit
comprising a timing signal generation circuit, a connection unit
for connecting the arithmetic elements, wherein the connection unit
inputs the pulse signals from the predetermined arithmetic elements
and executes a predetermined weighted time window integration, and
the arithmetic elements are disposed in parallel by the connection
unit and integrate the pulse modulation signals from the connection
unit on the basis of a timing signal from the timing signal
generation circuit.
Other objectives and advantages besides those discussed above shall
be apparent to those skilled in the art from the description of a
preferred embodiment of the invention which follows. In the
description, reference is made to accompanying drawings, which form
a part thereof, and which illustrates an example of the invention.
Such example, however, is not exhaustive of the various embodiments
of the invention, and therefore reference is made to the claims
which follow the description for determining the scope of the
invention.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 is a block diagram showing a whole architecture of a network
for a pattern detection/recognition system in one embodiment of the
present invention;
FIGS. 2A, 2B and 2C are diagrams showing configurations of a
synaptic portion and a neuron element portion;
FIGS. 3A and 3B are diagrams showing how a plurality of pulses are
propagated to feature detection layer neurons from a feature
integration layer or an input layer;
FIGS. 4A, 4B and 4C are diagrams showing an architecture of a
synaptic circuit;
FIGS. 5A, 5B and 5C are diagrams showing an architecture of a
synaptic connection small circuit, and an architecture of a pulse
phase delay circuit;
FIG. 6 is a diagram showing a network architecture when inputted to
the feature detection layer neuron from a pacemaker neuron;
FIGS. 7A, 7B, 7C, 7D and 7E are graphs showing a structure of a
time window, an example of a weighting function distribution and an
example of feature elements when processing a plurality of pulses
corresponding to the different feature elements, which are inputted
to feature detection neurons;
FIG. 8 is a diagram showing cells on respective layers;
FIGS. 9A and 9B are diagrams each showing an example of a neuron
configuration (array);
FIGS. 10A and 10B are diagrams showing a sharing structure of a
synaptic connection circuit;
FIGS. 11A and 11B are diagrams showing another sharing structure of
the synaptic connection circuit;
FIG. 12 is a diagram showing a further sharing structure of the
synaptic connection circuit;
FIG. 13 is a diagram showing a detailed architecture of the
synaptic connection circuit;
FIG. 14 is a timing chart showing behavioral timings of the
respective elements of the synaptic circuit in an embodiment 3;
FIGS. 15A and 15B are diagrams schematically showing a basic
architecture of the network in an embodiment 4;
FIG. 16 is a diagram schematically showing how weighting
coefficient signals are distributed to the respective synaptic
circuits from a weighting coefficient generation circuit; and
FIG. 17 is a diagram schematically showing an architecture of an
image input system mounted with an object recognition system using
a parallel pulse signal processing circuit in the embodiment 4.
DESCRIPTION OF THE PREFERRED EMBODIMENTS
First Embodiment
One preferred embodiment of the present invention will hereinafter
be described in detail with reference to the accompanying
drawings.
Outline of Whole Architecture
FIG. 1 is a diagram showing a whole architecture of a network for a
pattern detection/recognition system in the first embodiment. This
pattern detection/recognition system mainly deals with information
related to a recognition (detection) of an object or a geometrical
feature.
FIG. 1 illustrates a so-called convolutional network architecture
(LeCun, Y. and Bengio, Y., 1995, "Convolutional Networks for Images
Speed, and Time Series" in Handbook of Brain Theory and Neural
Networks (M. Arbib, Ed.), MIT Press, pp. 255 258). A different
point from the prior arts is, however, that inter-layer connections
within the same route can take a form of local interconnections
(which will hereinafter be described). A final output is defined as
a result of the recognition, i.e., a category of the object
recognized.
A data input layer 101 is a CMOS (Complementary Metal-Oxide
Semiconductor) sensor or a photoelectric converting device such as
a CCD (Charge Coupled Device) in the case of detecting and
recognizing an image, and a voice input sensor in the case of
detecting and recognizing a voice. Further, the data input layer 1
may be structured to input high-order data obtained from a result
of analysis (for example, a primary component analysis and so on)
by a predetermined data analyzing portion.
Given hereinafter is an explanation of the case of inputting the
image. A feature detection layer (1, 0) detects, based on a
multiple resolution process such as Gabor wavelet conversion and
others, a local low-order feature (that may contain a color
component feature in addition to the geometrical feature) of an
image patter by the number of a plurality of feature categories at
a plurality of scale levels or with a plurality of resolutions at
the same area in each of positions over the entire image surface
(or at each of predetermined sampling points over the entire image
surface). The feature detection layer (1, 0) is constructed of
neuron elements, each having a receptive field structure
corresponding to a category of feature amount (which is, e.g.,
gradients of line segments defined as a geometrical structure in
the case of extracting the line segments in a predetermined
direction as a geometrical feature), and generating a pulse train
corresponding to a degree thereof.
A feature detection layer (1, k) configures processing channels
with the plurality of resolutions (or at the plurality of scale
levels) on the whole (wherein k.gtoreq.0). Namely, when
exemplifying a case where the Gabor wavelet conversion is executed
on the feature detection layer (1, 0), a set of feature detection
cells with Gabor filter kernels having the same scale level but
different directional selectivities as a receptive field structure,
configure the same processing channel on the feature detection
layer (1, 0), and, on a subsequent layer (1, 1) also, the feature
detection cells (for detecting a higher-order feature) receiving
outputs from the former feature detection cells, belong to the same
channel as the above processing channel. On a still subsequent
layer (1, k) (wherein k>1) also, the feature detection cells
receiving the outputs from the plurality of feature integration
cells configuring the same channel similarly on a (2, k-1) layer,
are structured to belong to this channel. The processing at the
same scale level (or with the same resolution) proceeds through
each processing channel, wherein the low-order through high-order
features are detected and recognized by hierarchical parallel
processing.
A feature integration layer (2, 0) has a predetermined receptive
field structure (a receptive field 105 hereinafter implies a
connecting range with an output element of an immediate anterior
layer, and the receptive field structure implies a connecting
weight distribution), and is constructed of neuron elements each
generating the pulse train. The feature integration layer (2, 0)
integrates a plurality of neuron element outputs within the same
receptive field from the feature detection layer (1, 0) (the
integration involving an operation such as sub-sampling based on
local averaging, a maximum output extraction and so on). Further,
each receptive field of the neuron within the feature integration
layer has a structure common to the neurons within the same layer.
Each of the feature detection layers 102 (1, 1), (1, 2), . . . ,
(1, N) and the feature integration layers 103 ((2, 1), (2, 2), (2,
N)) has a predetermined receptive field structure. The former group
of feature detection layers ((1, 1), . . . ) detect, as by the
respective layers described above, a plurality of different
features in respective feature detection modules. The latter group
of feature integration layers ((2, 1), . . . ) integrate results of
detecting the plurality of features from the anterior feature
detection layers. The former group of feature detection layers are,
however, connected (wired) to receive cell element outputs of the
anterior feature integration layers belong to the same channel. The
sub-sampling defined as a process executed by the feature
integration layer involves averaging the outputs from local areas
(local receptive fields of the concerned feature integration layer
neurons) from a feature detection cell group coming under the same
feature category.
FIGS. 2A, 2B and 2C are diagrams showing a configuration of a
synaptic circuit and a configuration of the neuron element. A
structure for connecting inter-layer neuron elements 201 is, as
depicted in FIG. 2A, built by a signal transmission portion 203 (a
wire or a delay line) corresponding to an axon of a neural cell and
by synaptic circuits S202 corresponding to dendrites thereof. FIG.
2A shows the connecting architecture related to the outputs (that
are inputs if viewed from a certain feature detection (integration)
cell (N)) from a neuron group (n.sub.i) of a feature integration
(detection) cell that configures the receptive field with respect
to the feature detection (integration) cell (N). The signal
transmission portion 203 drawn by a bold line serves as a common
bus line, and pulse signals from the plurality of neurons, which
are arranged in time-series, are transmitted through on this signal
transmission line. The same architecture is also adopted in the
case of receiving the inputs from the cell (N) as an output
destination. In this case, the input signals and the output signals
may be processed in separation on the time-base absolutely in the
same architecture, or the processing may be executed in a way that
gives the same architecture as FIG. 2A shows in two systems for
inputting (the dendrite-side) and for outputting (the
axon-side).
The synaptic circuits S202 are categorized into those related to
the inter-layer connections (which are the connection between the
neurons on the feature detection layer 102 and the connection
between the neurons on the feature integration layer 103, and there
might exist the on-layer neuron connections to a posterior layer
and to an anterior layer), and those related to the connections
between the neurons within on the same layer. The latter type of
synaptic circuits are used, as the necessity may rise, mainly for
connections with pacemaker neurons that will be explained later on
and with the feature detection or integration neurons.
In the synaptic circuit S202, a so-called excitatory connection
involves amplifying the pulse signals, while an inhibitory
connection involves attenuating the signals. In the case of
transmitting the information through on the pulse signals, the
amplification and the attenuation can be actualized by any one of
an amplitude modulation, a pulse width modulation, a phase
modulation and a frequency modulation of the pulse signal.
According to the first embodiment, the synaptic circuit S202 is
used chiefly for a pulse phase modulation element, wherein the
signal amplification is converted into a substantial advance of a
pulse arrival time as a quantity intrinsic to a feature, whereas
the attenuation is converted into a substantial delay. Namely, the
synaptic connection, as will be mentioned later on, gives an
arrival position (phase) on the time-base that is intrinsic to the
feature in the neurons at the output destination, wherein the
excitatory connection gives a phase advance of the arrival pulse
with respect to a certain reference phase in terms of a qualitative
aspect, and the inhibitory connection gives a delay likewise.
Referring to FIG. 2A, each of neuron elements n.sub.j outputs the
pulse signals (a spiked signal train), and involves the use of a
so-called integrate-and-fire type neuron element as will be
explained below. Note that the synaptic circuit and the neuron
elements may, as illustrated in FIG. 2C, be combined to configure a
circuit block.
Neuron Element
Next, the neurons that form each layer will be described. Each of
the neuron elements is extension-modeled based on the so-called
integrate-and-fire type neuron, and is the same as this
integrate-and-fire type neuron in terms of such a point that the
neuron element fires when a result of linearly adding the input
signals (a pulse train corresponding to an action potential)
spatiotemporally exceeds a threshold value, and outputs the pulse
signals.
FIG. 2B shows one example of a basic architecture representing a
behavior principle of the pulse generation circuit (CMOS circuit)
defined as the neuron element, and illustrates what a known circuit
(IEEE Trans. On Neural Networks Vol. 10, p. 540) is extended.
Herein, the pulse generation circuit is constructed as what
receives the excitatory input and the inhibitory input.
The behavior principle of this pulse generation circuit will
hereinafter be explained. A time constant of a capacitor
C.sub.1/resistor R.sub.1 circuit on the side of the excitatory
input, is smaller than a time constant of a capacitor
C.sub.2/resistor R.sub.2 circuit on the side of the inhibitory
input. In a steady state, transistors T.sub.1, T.sub.2, T.sub.3 are
cut off. Note that the resistor is actually constructed of a
transistor connected in a diode mode.
When an electric potential of the capacitor C.sub.1 increases and
gets higher by a threshold value of the transistor T.sub.1 than
that of the capacitor C.sub.2, the transistor T.sub.1 becomes
active, and further the transistors T.sub.2, T.sub.3 get active.
The transistors T.sub.2, T.sub.3 form a current mirror circuit, and
an output of the circuit shown in FIG. 2B is given forth from the
side of the capacitor C.sub.1 by an unillustrated output circuit.
The circuit is structured so that when an electric charge
accumulation amount of the capacitor C.sub.2 is maximized, the
transistors T.sub.1 falls into a shutdown, then the transistors
T.sub.2, T.sub.3 are cut off as a result of the above shutdown, and
a positive feedback comes to 0.
During a so-called refractory period, the capacitor C.sub.2
discharges, and, if a potential of the capacitor C.sub.1 is larger
than a potential of the capacitor C.sub.2 and unless a difference
therebetween is over the threshold value of the transistor T.sub.1,
the neuron does not respond. The periodic pulses are outputted with
a repetition of alternate charging/discharging of the capacitors
C.sub.1, C.sub.2, and a frequency thereof is generally determined
corresponding to a level of the excitatory input. Owing to an
existence of the refractory period, the frequency can be, however,
restricted to the maximum value, and a fixed frequency can also be
outputted.
The potential, i.e., the charge accumulation amount of the
capacitor is controlled in terms of the time by a reference voltage
control circuit (time window weighting function generation circuit)
204. What reflects this control characteristic is a weighted
addition within a time window with respect to the input pulse,
which will be mentioned later on (see FIGS. 7A, 7B, 7C, 7D and 7E).
This reference voltage control circuit 204 generates a reference
voltage signal (corresponding to a weighting function in FIG. 7B)
on the basis of an input timing (or an interconnection input to the
neuron of the subsequent layer) from a pacemaker neuron that will
hereinafter be described.
The inhibitory input is not necessarily required in the first
embodiment in some cases, however, the input to the feature
detection layer neuron from the pacemaker neuron is set inhibitory,
whereby a divergence (saturation) of the output can be
prevented.
Generally, a relationship between the summation of the input
signals and the output level (the pulse phase, the pulse frequency,
the pulse width and so forth) changes depending on a sensitivity
characteristic of the neuron. This sensitivity characteristic can
be changed depending on a top-down input from a higher-order layer.
In the following discussion, it is assumed for an explanatory
convenience that circuit parameters be set so that a pulse output
frequency corresponding to the summation value of the input signals
rises steeply (therefore, the values are substantially binary in a
frequency domain) and that the output level (such as a timing with
a phase modulation added) and so on) fluctuates depending on the
pulse phase modulation.
Moreover, a pulse phase modulation portion may have an addition of
a circuit as shown in FIGS. 5A, 5B and 5C, which will hereinafter
be described. With this scheme, the weighting function in the time
window is controlled based on the reference voltage with the result
that the phase of the pulse output from this neuron changes, and
this phase can be used as an output level of the neuron.
A time .tau..sub.w1, as shown in FIG. 7B, corresponding to a
maximum value of the weighting function that gives a time
integrating characteristic (receiving sensitivity characteristic)
with respect to the pulse having undergone the pulse phase
modulation at the synaptic connection, is generally set earlier in
time than an arrival predicted time .tau..sub.s1 of the pulse
intrinsic to the feature given by the synaptic connection. As a
result, the pulse arriving earlier than the arrival predicted time
within a fixed range (the pulse arriving too early is attenuated in
the example in FIG. 7B) is, in the neuron receiving this pulse,
integrated in time as a pulse signal having a high output level. A
profile of the weighting function is not limited to a symmetry as
seen on Gaussian function etc and may assume an asymmetry. It
should be noted based on the gist elucidated above that the center
of each weighting function in FIG. 7B does not correspond to the
pulse arrival predicted time.
Further, an output phase of a (presynaptic) neuron has such an
output characteristic that a delay (phase from a fiducial time
corresponding to the beginning of the time window as will be
explained later on, is determined by the charge accumulation amount
after detecting phase synchronization when receiving the reference
pulse (based on the pacemaker output and others). A detailed
architecture of the circuit giving this output characteristic is
not essential to the present invention and is therefore omitted
herein. A pulse phase of a postsynaptic neuron is what the pulse
phase of the presynaptic neuron is added to an intrinsic phase
modulation amount given at the synapse concerned.
Further, there may also be utilized such a known circuit
architecture as to give forth an oscillatory output delayed by a
predetermined timing when the input summation value obtained by use
of the window function and so on exceeds the threshold value.
The architecture of the neuron elements using the neurons belonging
to the feature detection layer 102 or the feature integration layer
103, may take such a circuit architecture as to output the pulse
with a phase delay corresponding to the input level (the simple or
weighted summation value of the inputs described above) at which
the concerned neuron receives from the receptive field of the
anterior layer after receiving the pulse outputted from the
pacemaker neuron in a case where a firing pattern is controlled
based on an output timing of the pacemaker neuron that will be
mentioned later on. In this case, before the pulse signal from the
pacemaker neuron is inputted, there exists a transient state where
the respective neurons output the pulses in random phases with
respect to each other in accordance with the input levels.
The neuron of the feature detection layer 102 has, as explained
above, the receptive field structure corresponding to the feature
category, and outputs the pulse with an output level (given herein
in the form of the phase change; it may also be structured to show
a change based on the frequency, the amplitude and the pulse width)
taking a so-called squashing function value, i.e., such a
non-reductive and nonlinear function as to gradually saturate with
a fixed level, as in the case of, e.g., a sigmoidal function, etc.,
in accordance with a weight summation value (that will be explained
below) when this weight summation value depending on the time
window function of the input pulse signal from the neuron of the
anterior layer (the input layer 101 or the feature integration
layer 103) is equal to or larger than the threshold value.
Synaptic Circuit and Others
FIGS. 4A, 4B and 4C show a matrix layout of synaptic connection
small circuits each giving a synaptic connection strength (that
implies a magnitude of the modulation in regard to the phase delay
etc) to each of neurons n'.sub.j to which the neurons n.sub.i are
connected in the synaptic circuit 202 (S.sub.i)
As described above, each neuron of the feature detection layer 102
has the local receptive field structure (which is the local
synaptic connection structure to the anterior layer) in accordance
with the a structure of the pattern to be detected. This local
receptive field structure contains a plurality of synapses that
give a symmetry or a common connection strength. With respect to
the symmetry of the connection structure, excluding the symmetry of
a connection strength distribution pattern as viewed from the
neuron on the input side (a receiving side) of the signal, there
exists a symmetry (or a plurality of connection strengths taking
the same value) of a connection strength distribution pattern as
viewed from the neuron on the output side (a transmitting
side).
The former is typically a case where the feature detection layer
neuron for detecting a certain feature category inputs, partially
or totally with the same connection strength, pulse outputs
(strength levels) from the feature integration layer neurons
(corresponding to input layer pixels) with respect to a plurality
of different low-order feature categories (or input pixel
portions). For instance, each of the receptive field structures of
the feature detection layer neurons for performing the Gabor
wavelet conversion assumes the symmetry and has the same
sensitivity level (connection strength) in a plurality of positions
of the receptive fields.
The latter is a case where, for example, a plurality of pulse
signals having undergone an intrinsic modulation in the synaptic
connection element, are outputted to the plurality of neurons,
detecting the different feature categories, of the feature
detection layer defined as a subsequent layer from the feature
integration layer neuron representing a certain category, and a
synaptic connection pattern assumes a distribution symmetry (or
this connection pattern gives the same modulation amount in the
plurality of synaptic connections).
On the other hand, the connection to the neuron of the middle- or
high-order feature detection layer 102 from the low- or
middle-order feature integration layer 103 in the former structure,
generally can also take a non-local receptive field structure
(connection pattern) (see FIG. 1). If laid out as shown in FIG. 8
and FIGS. 9A and 9B, however, the local structure is obtained, and
the same symmetry (or the plurality of connection distribution
distributions at the same level within the same receptive field)
with respect to the local receptive field structure can be
given.
Namely, the non-local receptive field structure described above, in
the structure depicted in FIG. 1, the feature integration layer
neurons of a plurality of feature integration modules (indicated by
small rectangular areas on, e.g., a (2,0) layer in FIG. 1)
belonging to feature categories different from each other, are
arranged, even when related to a feature concerning the same
position on the input data, in positions that are spatially far
distant from each other if the feature integration module to which
these neurons belong differs, which means that the connection to
the feature detection layer from the feature integration layer
takes such a non-local wiring structure that the positional
proximity (or the coincidence) on the input data is not necessarily
the proximity in terms of wiring because of those outputs from the
plurality of feature integration layer neurons being inputted to
the feature detection layer neuron.
A structure shown in FIG. 8 and FIGS. 9A and 9B will hereinafter be
explained. The feature integration layer neurons with respect to
the plurality of geometrical features in a predetermined position
(or in local areas with this position being centered) on the input
data, are disposed adjacent to each other, and the respective
neurons have connections to the higher-order feature detection
later neurons. Referring to FIG. 8, a feature detection cell
F.sub.D (r, f.sub.k, i) is defined as a cell for detecting an i-th
feature category on a feature detection layer k in a position
corresponding to a location r on the input data. Further, a feature
integration cell F.sub.I (r, f.sub.k, i) is likewise defined as a
cell related to an i-th feature category on a feature integration
layer k in the position corresponding to the location r on the
input data. FIG. 8 schematically shows that each area showing,
together with the local receptive field, the inter-layer connection
to the feature integration layer from the feature detection layer
with respect to the low-order feature, has the local receptive
field structure in each inter-layer connection unlike FIG. 1.
For instance, if the number of the feature categories to be
extracted, is 4, the neuron elements corresponding to the
respective feature categories (F1, F2, F3, F4) are locally arrayed
in cluster as illustrated on the left side in FIG. 9A. Herein, the
neuron elements arrayed in cluster represents the feature
integration layer neurons with respect to the different geometrical
features in the same position on the input data. According to the
first embodiment, the feature integration layer involves the use of
an array structure as illustrated on the left side in FIG. 9A.
Further, FIG. 9A schematically shows wiring from the unspecified
neurons of the feature integration layer to a high-order
(corresponding to a middle level in the network as a whole) feature
detection layer.
On the feature detection layer, high-order feature categories
(which are herein two categories G1, G2) are detected in every
predetermined position on the input data. The neurons relative to
the category G1 receive outputs from the neurons (belonging to
elliptical areas defined by dotted lines in FIG. 9A and related to
an existence or non-existence of this category in positions
corresponding to a plurality of locations in the local area on the
input data) relative to F1, F2 of the feature integration layer.
Similarly, the neurons relative to the category G2 receive outputs
from the neurons related to F3, F4 of the feature integration
layer. Referring again to FIGS. 9A and 9B, the adjacent neurons of
the feature detection layer receive outputs from the neurons
(belonging to overlapped area segments of the ellipses in FIGS. 9A
and 9B) existing in overlapped area segments on the feature
integration layer.
FIG. 9B further schematically shows a part of the structure of
wiring from the feature detection layer to the feature integration
layer defined as a high-order layer of this feature detection
layer, and also an array of the neuron elements of the feature
integration layer corresponding thereto. Herein, feature categories
(g1, g2) of the respective neurons of the feature integration layer
are mapped to the feature categories (G1, G2), respectively, and
geometrical feature represents the same category (the
representations are distinguished from each other for the
convenience's sake). The respective feature integration layer
neurons receive the outputs from the plurality of neurons existing
in the local areas on the feature detection layer.
A circuit layout in matrix is that the signal lines of which a
synaptic connection strength (phase delay amount) is common are
clustered by a single synaptic small circuit with respect to every
common synaptic connection strength. To be specific, the input
signal lines from the plurality of different neurons are connected
to a shared synaptic small circuit that should exist on the input
side to the unspecified feature detection layer neurons, and
further the output signal lines to the plurality of different
neurons or the signal lines each to the single neuron are connected
thereto as the signal lines to an output destination (the feature
integration layer) from the concerned neurons.
Referring to FIGS. 10 and 11, there will be explained a sharing
process of the synaptic connection circuit, which is executed in a
case where a connection pattern (synaptic modulation distribution)
to the respective feature detection layer neurons assumes a
symmetry, or a case where a plurality of synaptic connections give
the same modulation amount.
In common throughout the respective Figures, feature detection
layer neurons M1, M2, M3 have connections in sequence through a
synaptic circuit (plus switch circuit) group, wherein one
connection is formed by a neuron group (N1 through N5) with N3
being centered on the feature integration layer, another connection
is formed by a neuron group (N3 through N7) with N5 being centered,
and a further connection is formed by a neuron group (N5 through
N9) with N7 being centered. The symbols D.sub.1, D.sub.2, D.sub.3
shown in the synaptic circuits (plus switch circuits) represent
pulse delay quantities at the respective synapses. Given herein by
way of a simple example is such a topology that in the connections
to the feature detection layer neurons, the output from the central
neuron receives the delay amount D.sub.2, the outputs from the
neurons most vicinal (adjacent on both sides) to the central neuron
receive D.sub.3, and the output from the neurons positioned at the
second distance from the central neuron.
Referring to FIG. 10A, the respective pulse signal outputs from the
plurality of neurons are given the fixed delays in the same
circuits and then outputted in branch to the different neurons in
accordance with the neurons on the input side by use of the synapse
plus branch circuits. The branch output implies that the pulses
modulated with the predetermined delay quantities are outputted to
the plurality of feature detection neurons, and, referring again to
FIG. 10A, the branch lines are wired as shown in FIG. 9B within the
synaptic circuit for giving the delay amount D.sub.1.
For example, the pulse output from the feature integration layer
neuron N2 is given the delay amount D.sub.1 in the synaptic circuit
and thereafter outputted in branch to only the feature detection
layer neuron M1. Herein, as shown in FIG. 9B, diodes are set on the
branch wires posterior to the delay element, whereby the outputs
from the specified neurons are outputted to the specified neurons
owing to the branch structure. The output from the neuron N4 is,
after being given the delay D.sub.1, outputted to the feature
detection layer neurons M1, M2. This is because the receptive
fields of the neurons M1, M2 are overlapped at N4. Further, the
pulse outputs from N5 are a signal outputted to M2 after receiving
the delay amount D.sub.2 and a signal outputted to M1 after
receiving the delay D.sub.3. The delays of the outputs from the
neurons and the detailed branch output structure are the same as
those shown in FIG. 11A.
FIG. 11A shows that a small circuit D.sub.ij in each synaptic
circuit gives a delay amount D.sub.i to the pulse and outputs it to
the neuron Mj. Further, the inputs to the respective synaptic
circuits giving the delay amount D.sub.i require a larger amount of
wires for connecting the small circuits different depending on the
branch output destinations than in FIGS. 10A and 10B.
As the input signal lines to the unspecified feature integration
layer neurons, in the case of implementing the local averaging of
uniform weighting for sub-sampling with respect to these neurons,
as will be explained later on, the synaptic circuit for giving the
predetermined pulse delays etc are not required to be provided
midway. In the case of executing a process such as non-uniform
local averaging etc, however, there may be taken the same
architecture of the synaptic connection circuit to the feature
detection layer neuron.
Each of those output signal lines may be connected to a
predetermined synaptic circuit or may also be connected as a simple
branch line (delay line or wire) to an output destination neuron.
Note that the signal, it is assumed, be a pulse signal inputted and
outputted in a voltage mode.
If the network takes such an architecture as to have a shared
connection mode (for representing the synaptic connection
structures of the plurality of neurons with the weighting
coefficient distribution given in 1) of the connection weights, the
delay amount (P.sub.ij given in the following formula (1)) at each
synapse can be uniformed within the same receptive field in some
cases unlike the case in FIGS. 3A and 3B. For example, the
connection to the feature integration layer from the feature
detection layer, if the feature integration layer performs
sub-sampling based on the local averaging (which is to be, however,
the uniform weighting) of the outputs of the feature detection
layer defined as an anterior layer thereto, may take the above
architecture without depending on the detection object (i.e.,
without depending on a category of the object).
In this case, as illustrated in FIG. 4C, a single circuit S.sub.k,i
suffices for forming each of the synaptic connection small circuits
401 in FIG. 4A, and this circuit architecture is particularly
economical. On the other hand, if the connection to the feature
detection layer from the feature integration layer (or a sensor
input layer) takes this circuit architecture, what the feature
detection neuron detects is such an event that the pulses
representing a plurality of different feature elements arrive
simultaneously (or arrive substantially at the same time).
As depicted in FIG. 4B, each of the synaptic connection small
circuits 401 is constructed of a learning circuit 402 and a phase
delay circuit 403. The learning circuit 402 adjusts the above delay
amount by changing a characteristic of the phase delay circuit 403.
Further, the learning circuit 402 stores a characteristic value
thereof (or a control value thereof) on a floating gate element or
on a capacitor connected to the floating gate element. The phase
delay circuit 403 is classified as a pulse phase modulation circuit
and is, as shown in FIG. 5A, configured by using, for instance,
monostable multivibrators 506, 507, resistors 501, 504, capacitors
503, 505 and a transistor 502. FIG. 5B shows respective timings of
a rectangular wave P1 ([1] in FIG. 5B) inputted to the monostable
multivibrator 506, a rectangular wave P2 ([2] in FIG. 5B) outputted
from the monostable multivibrator 506, and a rectangular wave P3
([3] in FIG. 5B) outputted from the monostable multivibrator
507.
Though a detailed explanation of an operational mechanism of the
phase delay circuit 403 is omitted, a pulse width of the
rectangular wave P1 is determined by a time till a voltage of the
capacitor 503 based on a charging current reaches a predetermined
threshold value, while a pulse width of the rectangular wave P2 is
determined by a time constant of the resistor 504 and the capacitor
505. If the pulse width of P2 expands (as indicated by a
dotted-line rectangular wave in FIG. 5B) and if a fall timing
thereof is shifted back, a rise time of P3 is shifted by the same
quantity, however, the pulse width of P3 remains unchanged, and it
therefore follows that the rectangular wave is outputted in a way
of being modulated by a phase of the input pulse.
A control voltage Ec is changed by the learning circuit 402 for
controlling the charge accumulation amount to a refresh circuit 509
having the reference voltage and to the capacitor 508 for giving
the connection weight, whereby the pulse phase (delay amount) can
be controlled. A long-term retainment of this connection weight may
involve storing the connection weight as charge of the floating
gate element (not shown) provided outside the circuit shown in FIG.
5A after the learning behavior or by writing it to a digital memory
and so on. There may be utilized other known circuit architectures
such as the architectures (refer to e.g., Japanese Patent
Application Laid-Open Nos. 5-37317 and 10-327054) each schemed to
downsize the circuit.
What is exemplified as the learning circuit at the synapse that
actualizes the simultaneous arrival of the pulses or the
predetermined phase modulation amount, includes the circuit
elements as shown in FIG. 5C. To be specific, the learning circuit
402 can be constructed of a pulse propagation time measuring
circuit 510 (a propagation time herein indicates a time difference
between a time of the pulse output of a presynaptic neuron on a
certain layer and an arrival time of this pulse at an output
destination neuron existing on a next layer), a time window
generation circuit 511, and a pulse phase modulation amount
adjusting circuit 512 for adjusting a pulse phase modulation amount
in the synaptic portion so that the propagation time takes a fixed
value.
The propagation time measuring circuit 510 involves the use of an
architecture for inputting clock pulses from the pacemaker neurons
configuring the same local receptive field as will be explained
later on and obtaining the propagation time based on an output from
a counter circuit for these clock pulses in duration of a
predetermined time width (time window: see FIG. 3B). Note that the
time window is set based on a point of firing time of the output
destination neuron, whereby Hebb's learning algorithm (rule)
extended as shown below is applied. Process (Extraction of
Low-Order Feature by Gabor Wavelet conversion etc) on Feature
Detection Layer (1, 0)
Supposing that the feature detection layer (1,0) contains the
neurons detecting a structure (low-order feature) of a pattern
having a predetermined spatial frequency in a local area having a
certain size and a directional component of being vertical and if
there exists a structure corresponding to an interior of the
receptive field of N1 on the data input layer 1, the neuron outputs
the pulse in phase corresponding to a contrast thereof. This type
of function can be actualized by a Gabor filter. A feature
detection filter function performed by each of the neurons of the
feature detection layer (1,0) will hereinafter be discussed.
It is assumed that the Gabor wavelet conversion expressed by a
filter set having multi-scales and multi-directional components on
the feature detection layer (1,0) and each of the intra-layer
neurons (or each group consisting of a plurality of neurons) has a
predetermined Gabor filtering function. On the feature detection
layer, one single channel is configured by clustering a plurality
of neurons groups each consisting of neurons having the receptive
field structures corresponding to a convolutional operation kernels
of a plurality of Gabor functions that have a fixed scale level
(resolution) and different directional selectivities. On this
occasion, the neuron group forming the same channel has a different
directional selectivity, and the neuron groups exhibiting the same
size selectivity may be disposed in positions adjacent to each
other, or the neuron groups belonging to different processing
channels may also be disposed adjacent to each other. This scheme
is based on an idea that the actualization is easier in terms of
the circuit architecture by adopting the layouts shown in the
respective Figures for the convenience's sake of a connecting
process that will be mentioned below in the group-oriented
coding.
Incidentally, for details of the method of executing the Gabor
wavelet conversion in the neural network, refer to a document (IEEE
Trans. On Acoustics, Speed, and Signal Processing, vol. 36, pp.
1169 1179) by Daugman (1988).
Each of the neurons of the feature detection layer (1,0) has the
receptive field structure corresponding to a kernel g.sub.mn. The
kernel g.sub.mn having the same scale index m has a receptive field
of the same size, and a corresponding kernel g.sub.mn size is set
corresponding to the scale index in terms of the operation. Herein,
the sizes such as 30.times.30, 15.times.15 and 7.times.7 are set on
the input image in sequence from the roughest scale. Each neuron
outputs the pulse at such an output level (which is herein on a
phase basis; an architecture on a frequency basis or an amplitude
basis or a pulse basis may also, however, be used) as to become a
nonlinear squashing function of a wavelet conversion coefficient
value obtained by inputting a sum of products of distribution
weighting coefficients and image data. As a result, it follows that
the Gabor wavelet conversion is executed as an output of this whole
layer (1,0).
Processes (Extractions of Middle- and High-Order Features) on
Feature Detection Layer
Unlike the feature detection layer (1,0), each of the neurons of
the subsequent feature detection layers ((1,1), (1,2), . . . )
forms, based on the so-called Hebb's learning algorithm etc, the
receptive field structure for detecting a feature intrinsic to a
pattern of a recognition object. On a more posterior layer, a size
of the local area in which to detect the feature becomes stepwise
more approximate to a size of the whole recognition object, and
geometrically a middle- or high-order feature is detected.
For instance, when detecting and recognizing a face, the middle-
(or high-order) feature represents a feature at
pattern-element-oriented levels such as eyes, a nose, a mouth etc
shaping the face. Between different channels, if at the same
hierarchical level (the same level in terms of a complexity of the
feature to be detected), a difference of the feature detected comes
under the same category but is what is detected by the scales
different from each other. For example, the (eye) defined as the
middle-order feature is detected as an (eye) having a different
size at a different processing channel. Namely, the scheme is that
the in-image (eye) having a given size is detected at the plurality
of processing channels exhibiting different scale level
selectivities.
Note that each of the neurons of the feature detection layer may
generally have such a mechanism as to receive, based on the output
of the anterior layer, an inhibitory (shunting inhibition)
connection in order to stabilize the output (without depending on
the extractions of the low- and high-order features).
Process on Feature Integration Layer
The neurons of the feature integration layers ((2,0), (2,1), . . .
) will be explained. As illustrated in FIG. 1, the connection to
the feature integration layer (e.g., (2,0)) from the feature
detection layer (e.g., (1,0)) is configured to receive inputs of
the excitatory connections from the neurons of the same category
(type) of feature elements of the anterior feature detection layer
within the receptive fields of the concerned feature integration
neurons. The function of the neuron of the integration layer is, as
explained above, the local averaging for every feature category,
the sub-sampling based on the maximum value detection, and so
on.
According to the former mode, the plurality of pulses of the same
category of feature are inputted, and then integrated and averaged
in the local area (receptive field) (alternatively, a
representative value such as a maximum value is calculated within
the receptive field), thereby making it possible to surely detect a
positional fluctuation and a deformation of the feature. Therefore,
the receptive field structure of the neuron of the feature
integration layer may be formed so as to become uniform (such as
being in a rectangular area having a predetermined size in any
cases and exhibiting a uniform distribution of the sensitivity or
the weighting coefficient therein) without depending on the feature
category.
Pulse Signal Processing on Feature Integration Layer
As discussed above, according to the first embodiment, the feature
integration cell is not structured to receive the timing control
from the pacemaker neuron on the feature detection layer with a
layer number (1,k) anterior thereto. The reason is that in the
feature integration cell, the neurons output the pulses in phase
(any one of the frequency, the pulse width and the amplitude may be
dependent, however, the phase is adopted in the first embodiment)
determined not by the arrival time pattern of the input pulse but
by, if anything, an input level (such as a temporal summation value
of the input pulses) within a fixed time range, and hence a time
window occurrence timing is not so important. Note that this does
not intend to exclude an architecture in which the feature
integration cell receives the timing control from the pacemaker
neuron on the anterior feature detection layer, and this
architecture is, as a matter of course, feasible.
Behavior Principle of Pattern Detection
Next, pulse encoding of a two-dimensional graphic pattern and a
detection method thereof will be explained. FIGS. 3A and 3B
schematically show how the pulse signals are propagated to the
feature detection layer from the feature integration layer (e.g.,
from the layer (2,0) to the layer (1,1) in FIG. 1). The neurons
n.sub.i on the side of the feature integration layer correspond to
feature amounts (or feature elements) different from each other,
while the neurons n'.sub.j on the side of the feature detection
layer get involved in detecting a higher-order feature (pattern
elements) obtained by combining the respective features within the
same receptive field.
An intrinsic delay (intrinsic to the feature) due to a pulse
propagation time and a time delay etc in the synaptic connection
(S.sub.j,i) to the neuron n'.sub.j from the neuron n.sub.i, occurs
in each inter-neuron connection. As a result, so far as the pulses
are outputted from the neurons of the feature integration layer,
pulses of a pulse train Pi are set to arrive at the neuron n'.sub.j
in a predetermined sequence (such as P.sub.4, P.sub.3, P.sub.2,
P.sub.1 in FIG. 3A), depending on a delay amount at the synaptic
connection that is determined by learning.
FIG. 3B shows a pulse propagation timing to a certain feature
detection cells (n'.sub.j) (detecting a higher-order feature) in a
layer having a layer number (1, k+1) from feature integration cells
n.sub.1, n.sub.2, n.sub.3 (individually representing different
categories of features) in a layer having a layer number (2, k) in
the case of executing the synchronization control of the time
window by using the timing signals from the pacemaker neurons that
will be mentioned later on.
Referring to FIG. 6, the pacemaker neurons n.sub.p which accompany
the feature detection neurons (n.sub.j, n.sub.k etc) for detecting
different categories of features, form the same receptive field as
that of the feature detection neurons and receives the excitatory
connection from the feature integration layer (or the input layer).
Then, the pulses are outputted to the feature detection neurons and
the feature integration neurons at a predetermined timing (or
frequency) determined by an input summation value (or an activity
level average value of the whole receptive field in order to
control so as to depend on a state of representing an action
characteristic intrinsic to the whole receptive field).
Further, the scheme in each feature detection neuron is that the
time windows are phase-locked to each other with its input serving
as a trigger signal but are not phase-locked before receiving the
input from the pacemaker neuron as described above, and each neuron
outputs the pulse with a random phase. Further, in the feature
detection neuron, a time window integration that will be explained
below is not performed before receiving the input from the
pacemaker neuron but is performed, which is triggered by the pulse
input from the pacemaker neuron.
Herein, the time window, which is determined for every feature
detection layer neuron (n'.sub.i), is common to the respective
neurons within the feature integration layer forming the same
receptive field with respect to the concerned cell and to the
pacemaker neuron, and gives a time range for a time window
integration.
The pacemaker neuron existing on the layer having a layer number
(1, k) (where k is a natural number) outputs the pulse output to
each feature integration cell of the layer having a layer number
(2, k-1) and the feature detection cell (the layer number (1, k) to
which the pacemaker neuron belongs, whereby the feature detection
cell gives a timing signal for generating the time window when
adding the inputs in time aspect. A start time of this time window
serves as a reference time for measuring an arrival time of the
pulse outputted from each feature integration cell. Namely, the
pacemaker neuron gives the timing for outputting the pulse from the
feature integration cell, and a reference pulse for a time window
integration in the feature detection cell.
Each pulse is given a predetermined quantity of phase delay when
passing via the synaptic circuit, and arrives at the feature
detection cell further via the signal transmission line such as the
common bus. A sequence of the pulse train on the time-base at this
time is expressed such as pulses (P.sub.1, P.sub.2, P.sub.3) drawn
by the dotted lines on the time-base of the feature detection
cell.
In the feature detection cell, if larger than the threshold value
as a result of the time window integration (normally the
integration is effected once; there may also be, however, executed
the electric charge accumulation involving the time window
integration effected multiple times or the averaging process
involving the time window integration effected multiple times) of
the respective pulses (P.sub.1, P.sub.2, P.sub.3), a pulse output
(P.sub.d) is outputted based on a termination time of the time
window. Note that the on-learning time window shown in the same
Figure is what is referred to when executing the learning algorithm
that will hereinafter be discussed.
Spatiotemporal Integration of Pulse Outputs and Network
Characteristic
Next, an arithmetic process of spatiotemporal weighting summation
(a weight summation) of the input pulses will be explained. As
shown in FIG. 7B, each neuron takes a weight summation of the input
pulses by use of a predetermined weighting function (e.g., Gaussian
function) for every sub time window (timeslot), and the summation
of weights is compared with a threshold value. The symbol
.tau..sub.j represents a central or peak position of the weighting
function of a sub time window j, and is expressed by a start time
reference (an elapse time since the start time) of the time window.
The weighting function is generally a function of a distance (a
deviation on the time-base) from a predetermined central position
(representing a pulse arrival time in the case of detecting a
detection target feature), and assumes a symmetry. Accordingly,
supposing that the central position .tau. of the weighting function
of each sub time window (timeslot) of the neuron corresponds to a
time delay after learning between the neurons, a neural network for
obtaining the spatiotemporal weighting summation (the weight
summation) of the input pulses can be defined as one category of a
radial basis function network (which will hereinafter be
abbreviated to RBF) in the time-base domain. A time window F.sub.Ti
of the neuron ni using Gaussian function as a weighting function is
given by:
.times..times..times..delta..function..tau..times..function..tau..sigma.
##EQU00001##
where .sigma. is a spread with respect to every sub time window,
and b.sub.ij is a coefficient factor.
Note that the weighting function may take a negative value. For
example, if a certain feature detection layer neuron is to detect
eventually a triangle and when detecting a feature (F.sub.faulse)
that is not apparently an element configuring this graphic pattern,
a connection from the feature detection (integration) cell and a
weighting function making a negative contribution can be given from
pulses corresponding to the concerned feature (F.sub.faulse) in the
summation value calculation process of the input so that the
detection of the triangle is not eventually outputted even if there
is a large contribution from other feature elements.
A spatiotemporal summation X.sub.i(t) of the input signals to the
neurons n.sub.i of the feature detection layer is given by:
.function..times..times..times..function..times..function..tau.
##EQU00002## where .epsilon..sub.j is an initial phase of the
output pulse from the neuron n.sub.j. If converged at 0 due to
synchronization firing with the neuron n.sub.i, or if the phase of
the time window is forcibly synchronized with 0 due to the timing
pulse input from the pacemaker neuron, .epsilon..sub.j may be set
to 0 at all times. When obtaining the weight summation on the basis
of the pulse input in FIG. 7A and the weighting function shown in
FIG. 7B, a time-varying transition of the weight summation value as
shown in FIG. 7E is obtained. The feature detection layer neuron
outputs the pulse when this weight summation value reaches a
threshold value (Vt).
The output pulse signal from the neuron n.sub.i is, as explained
above, outputted to the neuron of the high-order layer with a time
delay (phase) given by learning at such an output level as to
become a squashing nonlinear function of the spatiotemporal
summation (a so-called input summation) of the input signals
(wherein the pulse output takes a fixed frequency (binary) and is
outputted in a way that adds a phase modulation quantity serving as
the squashing nonlinear function with respect to the spatiotemporal
summation of the input signals to a phase corresponding to a fixed
delay amount determined by learning).
Learning Algorithm
The learning circuit 402 may be structured so that the time window
described above comes to have a narrower width as the frequency at
which an object having the same category is presented becomes
larger. With this contrivance, the learning circuit 402 behaves so
as to get close to a coincidence detection mode for detecting
simultaneous arrivals of the plurality of pulses as the pattern
category is more familiar (which means a larger presentation count
and a larger learning count). This scheme makes it possible to
reduce the time required for detecting the feature (to perform an
instantaneous detection behavior) but is unsuited to a fine
comparative analysis of the spatial layout of the feature elements
and to a distinction between the similar patterns and so forth.
In the learning process of the delay amount, for example, by
extending to a complex number domain, a complex connection weight
C.sub.ij between the feature detection layer neuron n.sub.i and the
feature integration layer neuron n.sub.j is given such as:
C.sub.ij=S.sub.ij exp(iP.sub.ij) (3) where the first i in the
function exp represents an imaginary number unit, S.sub.ij denotes
a connection strength, P.sub.ij indicates a phase. The phase
P.sub.ij is a phase corresponding to the time delay of the pulse
signal outputted to the neuron i from the neuron j at a
predetermined frequency. The connection strength S.sub.ij reflects
the receptive field structure of the neuron i, and has a structure
that generally differs corresponding to a recognition/detection
object. This is separately formed by learning (supervised learning
or self-organization), or is formed as a predetermined
structure.
On the other hand, the learning algorithm for the self-organization
relative to the delay amount is given by:
&.beta..times..times..times..times..times..times..pi..times..times..tau.
##EQU00003## where C is a time differential of C, .tau..sub.ij is
the time delay (a preset quantity) described above, and .beta.
(through 1) indicates a constant. When solving the above equation,
C.sub.ij converges at .beta. exp(-2.pi.i.tau..sub.ij), and hence
P.sub.ij converges at -.tau..sub.ij. Explaining an example of
applying the learning algorithm with reference to the on-learning
time window shown in FIG. 3B, only when both of presynaptic neurons
(n1, n2, n3) and postsynaptic neurons (feature detection cells)
fire in a time range of the learning time window, the connection
weight is updated based on the formula (4). Note that the feature
detection cells fire after an elapse of the time window in FIG. 3B
and may also fire before the elapse of the time window in FIG.
3B.
Process on Feature Detection Layer
Processes (for learning and recognition) executed mainly on the
feature detection layer will hereinafter be described. Each feature
detection layer inputs the pulse signals with respect to a
plurality of different features from the same receptive field
within the processing channel set at every scale level as explained
above, and calculates the spatiotemporal weighting summation (the
weight summation) and implements a threshold process. The pulse
corresponding to each feature amount arrives at a predetermined
time interval, depending on a delay amount (phase) predetermined by
learning.
Learning control of this pulse arrival time patter is not essential
to the first embodiment and is not therefore explained in depth.
For instance, however, to be brief, the pulse corresponding to the
feature element among the plurality of future elements configuring
a certain graphic patter, if most contributory to detecting this
pattern, arrives earlier, and, between the feature elements
showing, if intact, substantially the same pulse arrival time,
there is introduced a competitive learning scheme that the pulses
arrive away by a fixed quantity in time from each other.
Alternatively, there may be taken such a scheme that the pulses
arrive at time intervals different between predetermined feature
elements (configuring a recognition object and conceived important
in particular such as a feature exhibiting a large mean curvature,
a feature exhibiting a high rectilinearity and so forth).
According to the first embodiment, each of the neurons
corresponding to the respective low-order feature elements within
the same receptive field on a certain feature integration layer
defined as a anterior layer, synchronously fires (pulse output) in
a predetermined phase. Generally, there exist the connections to
the feature detection neurons, defined as the neurons of the
feature integration layer, for detecting, though different in their
positions, the same high-order feature (in this case, there are the
connections, configuring, though difference in their receptive
fields, the same high-order feature). At this time, as a matter of
course, the synchronous firing occurs also among these feature
detection neurons. The output level thereof (a phase basis is
herein taken; an architecture taking a frequency basis or an
amplitude basis or a pulse width basis may also, however, be
adopted) is, however, determined by a summation (or average etc) of
contributions from the plurality of pacemaker neurons that are each
given for every receptive field of the feature detection neuron. In
the interior of the time window, the weighting function has such a
profile that the peak value corresponds to a synaptic weight value.
Further, a means for actualizing the weighting addition within the
time window taking the peak value is not limited to the neuron
element circuit shown in FIGS. 2A, 2B and 2C and may be, as a
matter of course, actualized otherwise.
This time window corresponds more or less to a time zone excluding
the refractory period of the neuron. Namely, there is no output
from the neuron even by receiving whatever input during the
refractory period (a time range other than the time window),
however, the behavior that the neuron fires corresponding to the
input level in the time window excluding the time range, is similar
to that of the actual biological neuron. The refractory period
shown in FIG. 3B is a time zone from immediate after the firing of
the feature detection cell to a start time of the next time window.
A length of the refractory period and a width of the time window
can be, of course, arbitrarily set, and the refractory period may
not be set shorter than the time window as shown in FIG. 3B.
According to the first embodiment, as schematically shown in FIG.
6, the already-explained mechanism is that the start timing
described above is made common by means of inputting the timing
information (clock pulse) by the pacemaker neuron (pulse output at
a fixed frequency) receiving the inputs from the same receptive
field with respect to, for example, every feature detection layer
neuron.
If configured in this fashion, the synchronization control (even if
necessary) of the time window does not need effecting throughout
the network, and, even when the clock pulse fluctuates as described
above, the reliability of detecting the feature is not degraded
because of receiving uniformly an influence of the output from the
same local receptive field (the on-the-time-base positional
fluctuation of the window function becomes the same among the
neurons forming the same receptive field). A tolerance of scatter
in circuit element parameter also increases in order for the local
circuit control to enable the synchronization behavior with a
reliability to be attained.
For simplicity, the feature detection neuron for detecting the
triangle as a feature will be described. It is assumed that the
feature integration layer anterior thereto reacts to a graphical
feature (feature elements) such as L-shaped patterns (f.sub.11,
f.sub.12, . . . ) having multiple directions, combinational
patterns (f.sub.21, f.sub.22, . . . ) of line segments each having
a continuity (connectivity) to the L-shaped pattern and
combinations (f.sub.31, . . . ) of a part of two sides configuring
the triangle as depicted in FIG. 7C.
Further, f.sub.41, f.sub.42, f.sub.43 shown in FIG. 7C represent
features shaping the triangles having different directions and
corresponding to f.sub.11, f.sub.12, f.sub.13. The intrinsic delay
amount is set between the neurons forming the inter-layer
connection by learning, and, as a result of this, in the triangle
feature detection neuron, the pulses corresponding the principal
and different features shaping the triangle are set beforehand to
arrive at respective sub time windows (timeslots) (w.sub.1,
w.sub.2, . . . ) into which the time window is divided.
For instance, the pulses corresponding to combinations of the
feature sets each shaping the triangle on the whole as shown in
FIG. 7A, arrive first at the sub time windows w.sub.1, w.sub.2, . .
. , w.sub.n into which the time window is divided by "n". Herein,
the delay quantities are set by learning so that the L-shaped
patterns (f.sub.11, f.sub.12, f.sub.13) arrive at within w.sub.1,
w.sub.2, w.sub.3, respectively, and the pulses corresponding to the
feature elements (f.sub.11, f.sub.12, f.sub.13) arrive at within
w.sub.1, w.sub.2, w.sub.3, respectively.
The pulses corresponding to the feature elements (f.sub.31,
f.sub.32, f.sub.33) arrive in the same sequence. In the case shown
in FIG. 7A, the pulse corresponding to one feature element arrive
at the single sub time window (timeslot). The division into the sub
time windows has such a significance that an integration mode when
integrating those features, e.g., a processing mode such as setting
a condition that all the feature elements be detected or a
condition that a given proportion of features be detected and so
on, is to be enhanced in its changeability and adaptability by
individually surely detecting the pulses (detection of the feature
elements) corresponding to the different feature elements developed
and expressed on the time-base in the restive sub time windows.
For instance, under conditions where the recognition (detection)
object is a face and a search (detection) for an eye defined as one
of parts configuring the face is important (a case where the
priority of detecting the eye's pattern is set high in the visual
search), a reaction selectivity ((a detection sensitivity to a
specified feature) corresponding to a feature element patter
selectively configuring the eye can be enhanced by introducing a
feedback connection from a high-order feature detection layer. This
scheme makes it possible to detect the feature in a way that gives
a higher importance to a lower-order feature element shaping a
high-order feature element (pattern).
Further, assuming that the pulse corresponding to a more importance
feature is set previously to arrive at the earlier sub time window,
the feature exhibiting the higher importance is easier to detect by
setting a weighting function value in the concerned sub time window
larger than values in other sub time windows. This importance (the
detection priority among the features) is acquired by learning or
may also be predefined.
Accordingly, if on condition that there occurs an event such as
detecting a given proportion of feature elements, the division into
the sub time windows comes to have almost no meaning, and the
processing may be implemented in one single time window.
Note that the pulses corresponding to the plurality (three) of
different feature elements arrive respectively and may also be
added (FIG. 7D). Namely, it may be based on a premise that the
pulses corresponding to the plurality-of feature elements (FIG. 7D)
or an arbitrary number of feature elements, be inputted to one
single sub time window (timeslot). In this case, referring to FIG.
7D, the pulses corresponding to other feature elements f.sub.21,
f.sub.23 supporting the detection of an apex angle portion
f.sub.11, of the triangle, arrive at the first sub time window.
Similarly, the pulses corresponding to other feature elements
f.sub.22, f.sub.31 supporting the detection of an apex angle
portion f.sub.12 arrive at the second sub time window.
Note that the number of divisions into the sub time windows
(timeslots), the width of each sub time window (timeslot), the
feature class, and the allocation of the time intervals of the
pulses corresponding to the feature elements, are not limited to
those described above and can be, as a matter of course,
changed.
Second Embodiment
According to a second embodiment, as shown in FIG. 11B, a branch
circuit for outputting in branch the outputs from the synaptic
circuits is set as a characteristic component by use of a local
timing generation element (or pacemaker neuron) PN as shown in FIG.
11B. The branch circuit has a demultiplexer-wise function as will
be exemplified in an embodiment that follow. The branch circuit
outputs the output from each of the synaptic circuits to each of
the different feature detection layer neurons in accordance with
the timing signal transmitted from the local timing generation
element. This scheme enables farther wires from the synaptic
circuits to be more simplified than in the architecture in the
embodiment 1.
For instance, supposing that the timing generation element
generates a low-frequency pulse signal having a smaller pulse width
than a period of the pulse train, the branch circuit is structured
to output, to the feature integration layer neuron intrinsic to the
timeslot, the pulse signal from the synaptic circuit that is
inputted to within each of the timeslots obtained by equally
dividing a time width up to a next pulse rise from a fiducial time
corresponding to a rise time of each pulse signal by the number of
output destination neurons. For attaining this, the branch circuit
may suffice on condition that it has a switch element for
establishing a connection to a line different for, e.g., every
timeslot and functions as a demultiplexer (refer to a third
embodiment). Note that the process of the neuron receiving the
signal having undergone the modulation in the synaptic circuit, is
the same as in the embodiment 1.
Third Embodiment
A third embodiment involves the use of a circuit, as a synaptic
circuit, incorporating a demultiplexer function together that sets
variable a modulation amount at the delay element of the synaptic
circuit, controls the modulation (delay) amount of the pulse signal
on the basis of the timing signal from the local timing element and
branches by switching the output pulses. With this architecture
adopted, by contrast with the architecture in the embodiment
discussed above in which the synaptic circuit giving the different
delay amount is formed independently as the circuit that is at
least physically different, even such a synaptic circuit is shared
in a time-sharing manner, thereby scheming to further downsize the
circuit scale. Note that the process of the neuron receiving the
signal having undergone the modulation in the synaptic circuit is
the same as in the embodiment 1.
As shown in FIG. 12, a synaptic circuit S as a simplified version
of the circuit architecture in FIG. 11B provides branch output
destinations of the signals after being modulated and modulation
amounts (delay amounts) each different for every timeslot (see FIG.
8) obtaining by effecting time-sharing of a period up to a next
timing pulse on the basis of the fiducial time corresponding to a
rise time of the timing pulse from the local timing element.
The synaptic circuit is structured to include, as shown in FIG. 13,
a timeslot pulse generation circuit 1301, a delay amount modulation
circuit 1302, and a demultiplexer circuit 1303 constructed of a
switch array, an activation circuit etc.
The demultiplexer circuit 1303 has a characteristic function of
outputting, with a master clock pulse timing serving as a fiducial
time, inputted data pulse signals to different branch lines in a
predetermined sequence when a switch intrinsic to each of timeslots
(T.sub.1, T.sub.2, . . . , T.sub.n) is switched ON.
FIG. 14 is a timing chart schematically showing an example of
behaviors of the respective elements of the synaptic circuit S when
the demultiplexer circuit 1303 includes four sets of switch arrays
and four lengths of output lines. Referring to FIG. 14, respective
delay amounts (D.sub.1, D.sub.2, D.sub.3) correspond to the
modulation amounts in the respective synaptic circuits in the
preceding embodiment, and the pulse signal after being modulated is
outputted to the branch line as an output destination when each
switch comes to the ON-state.
The timeslot pulse generation circuit 1301, after inputting the
reference timing pulse from the local timing element described
above, generates the pulse signals at an interval of a
predetermined time that is shorter than a master cock interval in a
way that synchronizes the reference timing pulse as a master clock
(refer to the output of the timeslot pulse generation element in
FIG. 14), wherein a time width between the respective pulses
corresponds to the timeslot.
The delay amount modulation circuit 1302 includes a selector for
inputting the master clock from the local timing element described
above and selecting, based on the master clock, one of a plurality
of delay times preset in the input timing sequence of the pulse
signals from the timeslot pulse generation element. Namely, each of
the time-varying delay amounts has a certain magnitude quantized as
shown in FIG. 8, and the delay amount modulation circuit 1302
time-controls the delay amount by selecting one of the plurality of
preset fixed delay circuits for every timeslot. Then, the
demultiplexer circuit 1303 branch-outputs the pulse signals each
having undergone the delay modulation for every timeslot to the
neurons different from each other. Note that the delay amount may
be, as a matter of course, time-controlled by methods other than
the above-mentioned.
According to the embodiments discussed above, the pulse signal
processing circuit includes the modulation circuit for inputting
the plurality of pulsed signals from the different arithmetic
elements and modulating in common the predetermined signals among
the plurality of pulse signals, and the modulated pulse signals are
outputted in branch to the different signal lines. This
architecture yields such an effect that the modulation circuit,
which should provide the predetermined modulation amount to the
plurality of pulsed signals, is not required to be set for every
pulse (every input-side signal line) in the signal processing
system for transmitting the pulsed signals between the plurality of
arithmetic elements in a way that executes the predetermined
modulation on the pulsed signals, and the circuit scale can be
downsized.
Further, in the parallel processing circuit including the plurality
of neuron elements and the synaptic connection elements for
connecting the neuron elements, the synaptic connection element is
constructed by use of the pulse signal processing circuit, whereby
the parallel signal processing circuit using the pulse signals can
be simplified.
Moreover, in the pattern recognition system including the data
input unit of a predetermined dimension, the plurality of data
processing modules and the data output unit for outputting the
result of the pattern recognition, the data processing module has
the feature detection layer for detecting the plurality of features
and is constructed of the plurality of arithmetic elements
connected to each other by the predetermined synaptic connection
unit, the arithmetic elements within the data processing module
output the pulsed signals at the frequency or timing corresponding
to the arrival time pattern of the plurality of pulses within the
predetermined time window, the output unit outputs the result of
detection or recognition of the predetermined pattern on the basis
of the outputs from the respective arithmetic elements on the
processing layer, the synaptic connection unit includes the
modulation circuit for inputting the plurality of pulsed signals
from the different arithmetic elements and modulating in common the
predetermined signals among the respective pulses, the synaptic
connection unit outputting in branch the respective modulated
pulsed signals to the different signal lines, and the circuit scale
can be thereby downsized and simplified in the hierarchical signal
processing system for extracting the feature on the basis of the
spatiotemporal distribution of the pulsed signals.
Further, the parallel processing circuit has the structure in which
the plurality of synaptic connections are shared as one single
circuit on the basis of the distribution symmetry of the synaptic
connection to the predetermined neuron elements, thereby attaining
the reduction in the circuit scale in accordance with the degree of
symmetry of the synaptic connection distribution.
Further, the modulation circuit is structured as the delay circuit
for giving the predetermined delay to the pulse signal, whereby the
individual synaptic connection weight can be actualized by the
common pulse signal delay circuit.
Still further, the pulse signal processing circuit includes the
modulation circuit for inputting the plurality of pulsed signals
from the different arithmetic elements and effecting the
predetermined modulation on each pulse, and the branch circuit for
outputting in branch the modulated pulsed signals to the different
signal lines in the predetermined sequence in a way that gives the
predetermined delay to the pulsed signal. With this architecture,
the pulsed signals having undergone the common modulation can
arrive at the plurality of arithmetic elements (neurons) existing
at the spatially different pulse propagation distance from a
certain arithmetic element (neuron) at the same timing or the
timing based on the predetermined rule.
Fourth Embodiment
According to a fourth embodiment, in the convolutional network
architecture already explained above, a synaptic circuit block
implements a parallel pulse signal process including the time
window integration with respect to the output pulsed signals from
the respective feature integration layer neurons. According to the
embodiments discussed so far, the time window integration is
performed on the side of the neuron. By contrast, however, the time
window integration is executed in parallel on the side of the
synapse according to the fourth embodiment.
An architecture in FIG. 15A is that synaptic circuit blocks
(D.sub.1, . . . , D.sub.4) output the signals defined as results of
implementing the time window integration using the weighting
coefficient mapped to the synaptic weight value with respect to the
output pulses from the integration layer neurons (N.sub.1, . . . ,
N.sub.4), and the feature detection layer neuron (M.sub.1) adds the
signals after being integrated, thereby forming an internal state
of the neuron.
FIG. 15B shows a configuration of each of the synaptic circuits
D.sub.j. The synaptic circuit D.sub.j is structured to input a
pulse signal from the anterior hierarchy, a weighting coefficient
function signal and a timing signal for taking synchronization
between the layers. A time window integration circuit 1501 executes
the time window integration of an input pulse signal and a
weighting coefficient function. An output signal generation circuit
1502 generates and outputs a signal corresponding to a result of
the time window integration. Note that the output signal generation
circuit 1502 may output directly the result of the time window
integration.
A weighting coefficient value of the weighted time window
integration carried out in each of the synaptic circuits is given
as a function of time as in the first embodiment. What is
characteristic of the fourth embodiment is, however, that a peak
value thereof corresponds to S.sub.ij in the formula (4) in the
embodiment 1. Note that a profile of the weighting coefficient
function of the time window integration may be, as a matter of
course, set in the way described above.
The peak value of this weighting coefficient function is defined as
a synaptic connection weight value configuring the receptive field
structure needed for detecting that a local feature (which is,
e.g., a local feature as shown in FIG. 7C when the feature
detection layer neuron detects a triangle) at a predetermined
low-order level has a proper spatial geometrical relationship so as
to shape the feature detected by the feature detection layer
neuron.
The weighting coefficient function can be set corresponding to any
of a case where the output from the feature integration layer
neuron is a phase modulation pulse and a case where the same output
is a pulse width modulation signal. In the former case, however, a
profile of distribution of this value is, unlike the profile shown
in FIG. 7B, asymmetrical within each sub time window. In the latter
case, the weighting coefficient function takes a fixed value
corresponding to the synaptic connection weight value without
depending on the time.
For example, in the case of inputting the phase modulation pulse
signal, the weighting coefficient function linearly decreases
corresponding to a delay time from an arrival predetermined time if
a detection level for detecting the local feature concerned comes
to its maximum. In the case of the pulse phase modulation signal,
there is multiplied the weighting coefficient value of a level
corresponding to a phase delay (which is a delay from the reference
time, given by the timing signal, of the pulse arrival time)
corresponding to a magnitude of the output of the feature
integration layer neuron. For instance, a substantial product-sum
operation of the synaptic weight value and the neuron output is
performed by decreasing the weighting coefficient value linearly
corresponding to the delay of the pulse signal arrival time.
Further, when the pulse width modulation signal corresponding to
the neuron output is inputted to the synaptic circuit, a profile of
the weighting coefficient function may be set so that a result of
time integration of the time-varying weighting coefficient function
and the pulse width modulation signal is mapped to a result of
multiplication of the integration layer neuron output and the
synaptic connection weight value.
When executing the time window integration, normally the feature
detection layer neuron, the feature integration layer neuron
existing within the receptive field thereof and the synaptic
circuit, input and output the pulse signals in synchronization with
the predetermined timing signals. This timing signal may involve
the use of the signal from the pacemaker neuron as in the first
embodiment, or the clock signal supplied from outside.
A waveform of the weighting coefficient signal may be supplied from
outside by a function waveform generator, or alternatively a
digital circuit generates a digital waveform by LUT (Look-Up Table)
method and a function generation method, and thereafter the digital
waveform is converted by a D/A converter into an analog waveform
(Morie, et al., "Study and Research for AD Merged System Image
Feature Extraction LSI and Natural Image Recognition System Using
the Same LSI": Study of Specified Field in 2000, Mixed Integrated
Systems for Real-Time Intelligent Processing, pp. 41 47). Note that
it is easy to generate a high-precision voltage waveform serving as
a function of time.
As in the fourth embodiment, the product with the synaptic weight
can be calculated by the time window integration with respect to
the pulse modulation output signal from the neuron, whereby the
high-precision product-sum operation can be actualized with the
compact circuit architecture.
A weighting coefficient signal generation circuit 1601 distributes,
as shown in FIG. 16, an arbitrary analog non-linear function
(function of time) supplied from outside or generated with a high
precision within the chip as a weighting coefficient signal to each
synaptic connection circuit in the synaptic circuit block to the
feature detection layer neurons related to the same feature
class.
After undergoing the time window integration in the synaptic
circuit block, the respective signals (pulse signals) to be
outputted are outputted in branch to the predetermined neuron
circuits inside the neuron circuit block, wherein a summation is
taken. Referring again to FIG. 16, each of the synaptic circuits
arrayed line by line within the synaptic circuit block supplies
this weighting coefficient signal to every synaptic connection
circuit to the feature detection layer neuron, whereby an arbitrary
synaptic connection weight distribution within one receptive field
can be generated as a spatiotemporal function.
FIG. 17 illustrates an outline of an architecture of an image input
system (for instance, a camera, a video camera, a scanner and so
on) mounted with the pattern recognition system as an object
detection (recognition) system, which is based on the parallel
pulse signal process involving the elements in the architecture
shown in FIGS. 15A and 15B as basic elements.
Referring to FIG. 17, a photographic system 1701 includes an
imaging optical system 1702 containing a photographic lens and a
drive control mechanism for zoom photography, a CCD or CMOS image
sensor 1703, an imaging parameter measuring portion 1704, an image
signal processing circuit 1705, a storage portion 1706, a control
signal generation portion 1707 for generating control signals for
control of imaging conditions, a display 1708 serving as a
viewfinder such as EVF etc, a stroboscope light emitting portion
1709 and a storage medium 1710. Further, the photographic system
1701 further includes an object detection (recognition) system 1711
(the pattern recognition system constructed of the parallel pulse
signal processing circuit having the hierarchical structure in the
embodiments discussed above).
This object detection (recognition) system 1711 in this
photographic system 1701 detects (an existing position and a size
of), for example, a face image of a pre-registered figure from
within a picture photographed. When the position of this figure and
a piece of size data are inputted to the control signal generation
portion 1707, the control signal generation portion 1707 generates,
based on an output from the imaging parameter measuring portion
1704, control signals for optimally controlling a focus on this
figure, exposure conditions, a white balance and so on.
The pattern detection (recognition) system according to the present
invention is thus utilized for the photographic system, as a result
of which the detection of the figure etc and the optimal
photographic control (AF, AE etc) based on this detection can be
attained by actualizing the function of surely detecting
(recognizing) the object with a low consumption of electricity and
at a high speed (in real time).
According to the embodiments discussed so far, the pulse signal
processing circuit includes the modulation circuit for inputting
the plurality of pulsed signals from the different arithmetic
elements and modulating in common the predetermined signals among
the plurality of pulsed signals, and the modulated pulse signals
are outputted in branch to the different signal lines, wherein the
modulation circuit, which should provide the predetermined
modulation amount to the plurality of pulsed signals, is not
required to be set for every pulse (every input-side signal line)
in the signal processing system for transmitting the pulsed signals
between the plurality of arithmetic elements in a way that executes
the predetermined modulation on the pulsed signals, and the circuit
scale can be downsized.
Although the present invention has been described in its preferred
form with a certain degree of particularity, many apparently widely
different embodiments of the invention can be made without
departing from the spirit and the scope thereof. It is to be
understood that the invention is not limited to the specific
embodiments thereof except as defined in the appended claims.
* * * * *